Tropically planar graphs

Desmond Coles, Neelav Dutta, Sifan Jiang, Ralph Morrison, Andrew Scharf
August 14, 2019
arXiv:1908.04320v1 [math.AG] 12 Aug 2019

Abstract
We study tropically planar graphs, which are the graphs that appear in smooth tropical plane curves.
We develop necessary conditions for graphs to be tropically planar, and compute the number of tropically
planar graphs up to genus 7. We provide non-trivial upper and lower bound on the number of tropically
planar graphs, and prove that asymptotically 0% of connected trivalent planar graphs are tropically
planar.

1 Introduction
Tropical geometry studies discrete, combinatorial analogs of objects from algebraic geometry. In the case
of an algebraic plane curve, the tropical analog is a tropical plane curve, which has the structure of a one-
dimensional polyhedral complex, embedded in R2 in a balanced way. Each tropical plane curve has an
associated Newton polygon ∆, which is a lattice polygon. The curve is dual to a regular subdivision of ∆, as
discussed in [MS15]; we call the curve smooth if that subdivision is a unimodular triangulation. A smooth
tropical plane curve is illustrated in the middle of Figure 1, with its subdivided Newton polygon pictured
on the left.

Figure 1: A smooth tropical plane curve in the center, with its subdivided Newton polygon on the left and
its skeleton on the right

Inside of a smooth tropical plane curve is a graph called its skeleton, which is the largest subset onto
which the curve admits a deformation retract. If the polygon has g interior lattice points, then the skeleton
has genus (that is, first Betti number) equal to g. On the right side of Figure 1 we see the skeleton of the
smooth tropical curve; it has genus 4 because the Newton polygon has 4 interior lattice points. There is also
a natural way to assign lengths to the edges of the skeleton, making it a metric graph as in [Bak08]. In this
paper, we will only be concerned with the combinatorial structure of the graph and not the lengths.
Definition 1.1. A graph G is said to be tropically planar (or troplanar for short) if it is the skeleton of a
smooth tropical plane curve.
Since tropical plane curves are embedded in R2 , all tropically planar graphs are planar. They are also
connected and trivalent since they are dual to unimodular triangulations. In general, there is no known
efficient way to test if a given graph is troplanar, although an algorithm for finding all troplanar graphs of a
fixed genus g was designed and implemented in [BJMS15]. In fact, their algorithm went further: it found all
troplanar graphs, and determined which edge lengths were possible on those graphs inside of tropical plane

1
curves. For genus g = 2, 3, 4, and 5, they found that there are 2, 4, 13, and 38 troplanar graphs of genus g,
respectively.
In this paper we work to further our understanding of troplanar graphs. We start by developing certain
criteria that troplanar graphs must (or must not) satisfy, and by pushing the computations of [BJMS15]
further to determine the number of troplanar graphs of genus 6 and genus 7. We also determine asymptotic
upper and lower bounds on T (g), the number of troplanar graphs of genus g. Our upper bound in Theorem
11 √
4.14 shows that T (g) = O(2 3 g+O( g) ), which provides one of several proofs that as g → ∞, most connected
trivalent planar graphs are not tropically planar. Our lower bound in Corollary 5.5 shows that T (g) = Ω(γ g ),
where γ ≈ 2.47; this is an improvement on the best previously known result that T (g) = Ω(2g ). There is
still a wide gulf between these upper and lower bounds, which will hopefully be narrowed by future research.
Our paper is organized as follows. In Section 2 we provide necessary background on polygons, graphs,
tropical curves, and asymptotics notation. In Section 3 we present properties of troplanar graphs, reviewing
some from previous works as well as developing new ones; we also discuss our computations of troplanar
graphs of genus 6 and 7 through the lens of these results. In Section 4 we prove our upper bound on the
number of troplanar graphs of a given genus, and in Section 5 we prove our lower bound.
Acknowledgements. The authors are grateful for their support from the 2017 SMALL REU at Williams
College, and from the NSF visa NSF Grant DMS1659037.

2 Background and definitions

In this section we establish background necessary for stating and proving our results. This material will
cover background on lattice polygons, graphs, and tropical curves, as well as some notation.

2.1 Lattice polygons

Any point in R2 with integer coordinates is called a lattice point. A line segment with lattice endpoints has
lattice length equal to 1 less than the number of lattice points on it. A lattice polygon is a polygon whose
vertices are lattice points. Unless otherwise stated, all polygons we consider will be lattice polygons, and will
be convex. Let ∆ be a lattice polygon with r boundary lattice points and g interior lattice points. We refer
to g as the genus of the polygon. It turns out that the numbers g and r encode a great deal of information
about the polygon, as illustrated in the following result.
Theorem 2.1 (Pick’s Theorem). Let ∆ be a lattice polygon with g interior lattice points r boundary lattice
points. Then the area of ∆ is given by
r
+ g − 1.
2
We say two lattice polygons ∆ and ∆0 are equivalent if one is obtained from the other by a matrix
transformation A ∈ PSL2 (Z). It turns out that if we fix g ≥ 1, there are only finitely many polygons of
genus g, up to equivalence. See [BJMS15, Proposition 2.3] for a discussion of this fact, and see [Cas12] for an
algorithm to compute all polygons of genus g. The lattice width of a polygon ∆ is the width of the smallest
horizontal strip containing some polygon ∆0 equivalent to ∆.
Let ∆(1) be the convex hull of the g interior lattice points of ∆. We refer to ∆(1) as the interior polygon
of ∆. If ∆(1) is a two-dimensional polygon, we say ∆ is nonhyperelliptic; otherwise, we say ∆ is hyperelliptic.
Thus for a hyperelliptic polygon, ∆(1) is either the empty set, a single point, or a line segment. Three polygons
of genus 3 are illustrated in Figure 2; the first is hyperelliptic, and the other two are nonhyperelliptic.

Figure 2: Three lattice polygons; the first is hyperelliptic, and the first two are maximal

2
 We say that a lattice polygon ∆ is maximal if it is maximal with respect to containment among all
polygons with interior polygon ∆(1) , i.e. if there exists no lattice polygon ∆0 properly containing ∆ with
(∆0 )(1) = ∆(1) . The first two polygons in Figure 2 are maximal, while the third is not. An important tool
when studying maximal nonhyperelliptic polygons is [Koe91, Lemma 2.2.13], which states that any maximal
nonhyperelliptic polygon ∆ is obtained by “moving out” the edges of its interior polygon ∆(1) . For instance,
the middle polygon in Figure 2 can be obtained by moving out the edges of its interior lattice triangle. It
follows that given any two-dimensional lattice polygon Σ, either there exists no lattice polygon ∆ whatsoever
with ∆(1) = Σ, or there exists a unique maximal lattice polygon ∆ with ∆(1) = Σ. See [BJMS15, §2] for
more discussion.
We now move on to subdivisions of lattice polygons. A subdivision of a lattice polygon is a partition of
that polygon into lattice subpolygons, such that the intersection of any two subpolygons is a mutual face
(either an edge, a vertex, or the empty set). A triangulation is a subdivision where each subpolygon is a
triangle. A triangulation is called unimodular if each triangle has the minimum possible area, which by
Pick’s Theorem is 21 . We sometimes call a lattice triangle of area 12 a unimodular triangle. One way to
construct a subdivision of ∆ is to assign values aij ∈ R to each lattice point (i, j) in ∆, and to take the
convex hull of the points (i, j, aij ) in R3 . The assignment of aij ’s to the lattice points is called a height
function. We then project the lower faces of this polyhedron onto ∆, giving us a subdivision. This process is
illustrated on the left in Figure 3. Any subdivision that arises from a height function is called regular. Thus
the left triangulation in Figure 3 is regular; it turns out that the right triangulation is not regular [DLRS10,
Example 2.2.5]. A split in a subdivision is an edge of lattice length one with endpoints on different boundary
edges of the lattice polygon. Any split divides a polygon into two lattice polygons. If both these lattice
polygons have positive genus, we call the split nontrivial. For instance, the subdivision in Figure 1 has a
nontrivial split, separating the triangle into a triangle of genus 1 and a quadrilateral of genus 3.

Figure 3: The process of inducing a regular triangulation with a height function; and a nonregular triangu-
lation on the right

2.2 Graphs
A graph G = (V, E) is a finite collection of vertices V joined by a finite collection of edges E. We allow
multiple edges between a pair of vertices, and we also allow loops, which are edges from a vertex to itself.
We call a graph connected if it is possible to move from any vertex to any other vertex using the edges. We
say a graph is planar if it can be drawn in R2 without any edges crossing each other.
The degree of a vertex is the total number of edges incident to that vertex, where each loop is counted
twice. If a vertex has degree n, we will refer to it as an n-valent vertex. We say that a graph is trivalent
if every vertex has degree 3. An edge e in a graph G is called a bridge if the graph G \ {e} obtained by
deleting e from G has more connected components than G. If a graph is connected and has no bridges, we
call G bridgeless, or equivalently 2-edge-connected. The connected components that remain after deleting all
bridges from a connected graph G are called the 2-edge-connected components of G.
An isomorphism from G = (V, E) to G0 = (V 0 , E 0 ) is a bijection ϕ : V → V 0 such that the number
of edges between v, w ∈ V is equal to the number of edges between ϕ(v), ϕ(w) ∈ V 0 . If there exists an

3
isomorphism from G to G0 , we say that G and G0 are isomorphic. Virtually every property of a graph is
preserved under isomorphism, including the number of bridges and the structure of the 2-edge-connected
components.
The genus of a connected graph is g(G) := |E| − |V | + 1; this is also known as the first Betti number of
the graph. By Euler’s Polyhedron Formula, if G is planar then g(G) is the number of bounded regions in any
planar drawing of G. The graphs we are most concerned with in this paper are those that are connected and
trivalent, with genus g ≥ 2. We denote the number of such graphs of genus g as G (g), and the number of such
graphs that are planar as P(g). The graphs of genus 2 and 3 are illustrated in Figure 4, so G (2) = P(2) = 2
and G (3) = P(3) = 5. It turns out there is a single nonplanar connected trivalent graph of genus 4, namely
the complete bipartite graph K3,3 , so G (4) = P(4) + 1. The connected trivalent graphs were enumerated
up to genus 6 in [Bal76], which found there to be 2, 5, 17, 71, and 388 such graphs of genus 2, 3, 4, 5, and 6,
respectively. We remark that the literature does not always use the term genus, and instead stratifies these
graphs by the number of vertices; there is no harm in this, since for any connected trivalent graph we have
|V | = 2g − 2.

Figure 4: The connected trivalent graphs of genus 2 and 3

2.3 Tropical curves and their skeletons

We now briefly review the tropical geometry necessary for our paper; see [MS15] for more details. Tropical
plane curves are defined by polynomials p(x, y) over the tropical semiring (R ∪ {∞}, ⊕, ), where a ⊕ b =
min{a, b} and a  b = a + b. The subset of R2 defined by p(x, y) is the set of points where the minimum
in the polynomial is achieved at least twice. By the Structure Theorem [MS15, Theorem 3.3.5], a tropical
plane curve is a balanced 1-dimensional polyhedral complex, consisting of edges and rays meeting at vertices.
Forgetting about the embedding into R2 , this means we can interpret a tropical curve as a graph with a
1-valent vertex at the end of each of the rays.
The Newton polygon of a tropical polynomial p(x, y) is the convex hull of all exponent vectors of terms
that appear in p(x, y) with non-∞ coefficients. By [MS15, Proposition 3.1.6], every tropical plane curve
is dual to a regular subdivision of its Newton polygon; in particular, to the regular subdivision induced
by the coefficients of the polynomial. (It follows that every regular subdivision of a lattice polygon has a
tropical curve dual to it.) This duality means that a tropical curves has one vertex for each subpolygon in
the subdivision; that two vertices are joined by an edge if and only if the dual subpolygons share an edge;
and that there is a ray for each boundary edge of a subpolygon.
A tropical plane curve is called smooth if the corresponding subdivision of its Newton polygon is a
unimodular triangulation. Each (finite) vertex is then incident to a total of three rays and vertices. A
smooth tropical plane curve has first Betti number equal to the genus of its Newton polygon.
Now assume that the Newton polygon of tropical curve has genus g ≥ 2. The skeleton of a smooth
tropical plane curve is the graph that is obtained by removing all rays; iteratively retracting any leaves and
their edges; and then smoothing over any 2-valent vertices. This skeleton is a connected trivalent planar
graph of genus g. As defined in the introduction, any graph that is the skeleton of some smooth tropical
plane curve is called tropically planar, or simply troplanar. Six smooth tropical plane curves are illustrated
in Figure 5, along with skeletons below and dual Newton subdivisions above. From these examples, we know
that the first six graphs in Figure 4 are troplanar; it follows from Proposition 3.1 that the seventh graph is
not.
The authors of [BJMS15] implemented an algorithm for computing all troplanar graphs of a fixed genus,
including the achievable edge lengths. Ignoring this edge length computation, their algorithm can be sum-
marized as follows:

4
 Figure 5: Newton subdivisions, tropical curves, and skeletons

(1) Find all lattice polygons with g interior lattice points, up to equivalence.
(2) Find all regular, unimodular triangulations of each polygon P found in step (1).
(3) For each triangulation found in step (2), compute the dual graph, and find the skeleton.

It is this simplified version of the algorithm that we apply in Section 3 to find the number of troplanar graphs
of genus 6 and genus 7. We also use a key observation made by the authors of [BJMS15], namely that it
suffices in step (1) to consider maximal polygons: it follows from [BJMS15, Lemma 2.6] that any graph that
arises from a nonmaximal polygon of genus g will also arise from a maximal polygon of genus g, namely
from any maximal polygon containing the original polygon and having the same interior lattice points.

2.4 Asymptotics notation

We close this section by briefly recalling big O, little O, and big Omega notation. Let f (x) and g(x) be
functions from R to R (although similar notation will hold if the domain is N). We write f (x) = O(g(x))
as x → ∞ if for all sufficiently large values of x, the absolute value of f (x) is at most a positive constant
multiple of g(x). We write f (x) = o(g(x)) as x → ∞ if for all ε > 0, there exists x0 such that for x > x0 we
have f (x) ≤ εg(x). There are many different conventions for big Omega notation; for the purposes of this
paper, we write f (x) = Ω(g(x)) as x → ∞ if g(x) = O(f (x)) as x → ∞. When clear from context, we omit
the “as x → ∞” from all these notations. We write f (x) ∼ g(x) when limx→∞ fg(x) (x)
= 1.

3 Properties of troplanar graphs

In this section we will discuss some properties and invariants of troplanar graphs. As already noted, they are
connected trivalent planar graphs. In general, they are a difficult family of graphs to study. For example, the
set of troplanar graphs is not minor closed: the graph of genus 4 from Figure 6 is troplanar, as demonstrated
by the pictured smooth tropical curve containing it as a skeleton; but the minor obtained from collapsing
the central cycle is not troplanar, due to Proposition 3.1 below.
We review three previously known criteria for deducing that a graph fails to be troplanar. First, if a
graph is nonplanar, then it is not troplanar. We say that a connected graph is sprawling if it has a vertex
such that deleting the vertex from the graph creates three or more components. We say a planar embedding
of a graph is crowded if either two faces share two or more edges with one another, or one face shares an
edge with itself. If a graph is planar such that all its planar embeddings are crowded, we say that the graph
itself is crowded.
Proposition 3.1 (Proposition 4.1 in [CDMY16]). If G is sprawling, then G is not troplanar.

5
 Figure 6: A troplanar graph of genus 4, with a non-troplanar minor of genus 3

The original proof of this property uses a balancing property of tropical curves. An alternate proof
presented in [BJMS15] proves that the required dual structure in a Newton polygon cannot arise in a
unimodular triangulation.
Proposition 3.2 (Lemma 3.5 in [Mor17]). If G is crowded, then G is not troplanar.
This result follows readily from the fact that no graph dual to a triangulation can have two faces sharing
two edges, or a face that shares an edge with itself. It is not always immediately obvious is a graph if
crowded, since we need information about all planar embeddings of that graph; see [Mor17, §3] for methods
of checking whether a graph is crowded.
For the 17 distinct genus 4 connected trivalent graphs, these properties (nonplanar, sprawling, and
crowded) are enough to rule out all non-troplanar graphs: 13 of the graphs are troplanar [BJMS15, §7],
while 1 is nonplanar, and 3 are sprawling. (It turns out that no genus 4 graph is crowded.)

Figure 7: Three genus 5 graphs easily shown not to be troplanar

Figure 7 illustrates several graphs of genus 5. The first is nonplanar, the second is sprawling, and the
third is crowded by [Mor17, Example 3.4]. Thus, all three are readily determined by our criteria not to
be troplanar. Unfortunately, the results presented thus far are not enough for all graphs of genus 5: as
computed in [BJMS15], the seven graphs in Figure 8 are not troplanar, even though none are nonplanar,
sprawling, or crowded.
We now present new results to help us further determine which graphs are troplanar and which are not.
The following criterion is similar to the sprawling criterion in that it provides a structure that is forbidden
for troplanar graphs.
Definition 3.3. A connected trivalent planar graph is called a TIE-fighter graph if it has the form illustrated
in Figure 9, where each shaded region represent a subgraph of positive genus.

Theorem 3.4. Any TIE-fighter graph is not troplanar.

Proof. Suppose for the sake of contradiction that there exists a TIE-fighter graph G that is troplanar. Let ∆
be a convex lattice polygon with T a unimodular triangulation of ∆ such that the corresponding troplanar
graph is a TIE-fighter graph. The planar embedding of G coming from T must have C as a bounded face
with the bridges incident to v1 and v2 exterior to it: otherwise the embedding would be crowded. The face
formed by C is thus dual to some lattice point p of ∆. Then there must two unimodular triangles, t1 , and
t2 , where ti is dual to the vertex vi in the bridge ei connected to Hi . Since v1 and v2 lie on the cycle C in
G, the triangles t1 and t2 in T must intersect in a shared vertex, namely the lattice point p dual to the face
C. Moreover, and the edges of t1 and t2 not intersecting must be nontrivial splits of ∆, since they yield the

6
 Figure 8: Seven genus 5 graphs that are not troplanar

G1

v1 v2
H1 C H2

G2

Figure 9: The form of a TIE-fighter graph

bridges e1 and e2 . Without loss of generality let p be at the origin. Let these splits be s1 and s2 (so that si
is an edge of ti ), and let li be the line passing through si .
We now split into two cases, both illustrated in Figure 10. First, assume l1 and l2 are parallel. After
a change of coordinates we may assume both l1 and l2 are vertical. Thus any lattice point in s1 has x-
coordinate -1, and any lattice point in s2 has x-coordinate 1; otherwise t1 and t2 would not have area 1/2
each. Note that ∆ must contain the lattice points (0, 1) and (0, −1), because G1 and G2 have genus greater
than zero, and any lattice point in ∆ corresponding to a bounded face from some Gi must be contained in
the same connected component of ∆ \ (s1 ∪ s2 ) as the origin. Now let the lattice point on s2 with the lower
y-coordinate be (1, y); perhaps applying a reflection over the x-axis, we may assume y ≥ 0. Because ∆ is
convex and s2 is a split, we now observe that any lattice point dual to some cycle of H2 must be strictly
contained in the region bounded by the lines k1 and k2 , where k1 passes through (1, y) and the origin,
and k2 passes through (0, 1) and (1, y + 1). But there are no lattice points in the interior of this region, a
contradiction.
For the second case, assume l1 and l2 are not parallel. Let P be the intersection point of l1 and l2 .
Without loss of generality we may assume P lies above the x-axis. Let Pi be the lattice point in si nearest
P . Since G1 and G2 both have genus greater than 0, there must be some interior lattice point Q of ∆
contained in the interior of the convex hull of p = (0, 0), P1 , P2 , and P . We have that Q is closer to the
intersection point P than p is, and it follows that Q has Euclidean distance to either l1 or l2 strictly smaller
than that of p to that line. This means that for some i, the lattice triangle formed by the convex hull of si
and Q has strictly smaller area than ti . However, this is impossible since ti has the minimum possible area
of 1/2, a contradiction.
Having reached a contradiction is both cases, we conclude that any TIE-fighter graph is not troplanar.
The first three graphs in Figure 8 are all TIE-fighter graphs, so Theorem 3.4 provides a proof that they
are not troplanar. We also immediately obtain the following corollary, which forbids a graph from having

7
 P
k2

l1 k1
s2 Q
(1, y) P1 P2
s2
(0, 0) s1
p

l2
l1 l2

Figure 10: Illustrations for the two cases of our proof

too many loops in a row.

Corollary 3.5. Let G be a troplanar graph with of genus g ≥ 5. Then G cannot have three vertices on a
path, each incident to an edge that is incident to a loop.
Proof. Suppose G is a troplanar graph with three vertices in a path, each incident to an edge that is incident
to a loop. This path must be part of a cycle: otherwise the graph would be sprawling, as removing the
middle vertex would disconnect the graph into three components. Thus, G must have the form illustrated
on the left in Figure 11. If the boxed area has positive genus, then G is a TIE-fighter graph as illustrated on
the right of the figure, and so cannot be troplanar. Thus the boxed area must have genus 0, meaning that
G is the genus 4 graph from Figure 6. In particular, we cannot have g ≥ 5.

=
G0
G0

Figure 11: A graph that is a TIE-fighter if g ≥ 5

We now present two more results, both involving bridges, that relate the troplanarity of different graphs
to one another.
Proposition 3.6. Let G be a troplanar graph with a bridge e. Then the connected components of G \ {e},
after smoothing over 2-valent vertices, are troplanar.
By convention, if one of the components is simply a loop, we consider that loop to be troplanar.
Proof. Let G be a troplanar graph with a bridge e, and let ∆ be a Newton polygon with a regular unimodular
triangulation T giving rise to G. The bridge in a troplanar graph G corresponds to a split T . The split
subdivides ∆ into two sub polygons, ∆1 and ∆2 . Since the starting triangulation was regular, the resulting
triangulations T1 and T2 of ∆1 and ∆2 are regular as well: the height function that induced the triangulation

8
on ∆ can be restricted to each ∆i to give the restricted triangulations. The connected components G1 and
G2 of G \ {e} are then troplanar since they arise from T1 and T2 . This is illustrated in Figure 3.

Figure 12: Splitting a polygon into smaller polygons, and splitting a troplanar graph into smaller troplanar
graphs

This allows us to construct families of non-troplanar graphs of arbitrarily high genus that are not non-
planar, sprawling, crowded, or TIE-fighters graphs. In particular, we can take one of the four graphs on
the bottom of Figure 8, add a bridge e on any edge that is not already a bridge, and attach any troplanar
graph of any genus. This graph cannot be troplanar, since removing e yields two graphs, one of which is not
troplanar.
We now present a graph theoretic surgery on bridges for preserving troplanarity. This move is inspired
by the well-known bistellar flip in a triangulation. Unfortunately bistellar flips do not in general preserve
the regularity of triangulations, but we make use of them in a case where they do.
Let b be a bridge of a troplanar graph G with end vertices v and w, as illustrated in Figure 3. Let the
other two edges (possibly non-distinct in the case of a loop) emanating from v be e1 , and e2 , and let f1 ,
and f2 be the other edges emanating from w. Delete b, so that v and w are both 2-valent. Split v into two
distinct univalent vertices v1 and v2 , with vi a 1-valent vertex incident to ei . Do the same for w. Identify v1
with w1 to make a new vertex v 0 , and v2 with w2 to make w0 , and add an edge b0 between v 0 and w0 . This
process is illustrated in Figure 3. We say the new graph is the bridge reduction of G with respect to b.

e1 f1 e1 f1
e1 f1 e1 f1 e1 f1

v1 w1 v0 v0
v w v w b0
b
v2 w2 w0 w0

e2 f2 e2 f2 e2 f2
e2 f2 e2 f2

Figure 13: Bridge reduction

Proposition 3.7. Let G be a troplanar graph with a bridge b. Then the bridge reduction of G with respect
to b is troplanar.
Proof. For a height function ω, write T (ω) for the triangulation induced by ω. Let ∆ be a lattice polygon
and ω a height function such that the triangulation T (ω) is a unimodular triangulation of ∆ giving rise to
G. The bridge in G is dual to a split in T (ω). Let a and b be the endpoints of the split, and let c and d
be the other points in the two triangles containing the split. We will refer to the quadrilateral containing
these triangles be abcd. Since a and b form a split and since ∆ is convex, we know that abcd is convex.
The surgery for bridge reduction corresponds to a bistellar flip in abcd; that is, it corresponds to removing
the split from a to b, subdivide abcd with a segment from c to d. This is illustrated in Figure 14. All that
remains to show is that the triangulation obtained from this bistellar flip is still regular.
The split between a and b subdivides ∆ into two polygons, ∆1 and ∆2 . Let ω1 = ω|∆1 and ω2 = ω|∆2 be
the restrictions of ω to these polygons. Note that the triangulation T (ωi ) is simply the triangulation T (ω)
restricted to ∆i .

9
 a a

c d −→ c d

b b

Figure 14: A bistellar flip

T (ω) T (ω1 ) = T (ω10 ) T (ω2 ) = T (ω20 ) T (ω 0 ) T (ω 00 )

c d c d c d c d

∆ ∆1 ∆2 ∆ ∆
Figure 15: The starting triangulation of ∆; the restricted triangulations of ∆1 and ∆2 ; and the next two
triangulations of ∆

By [DLRS10], we may choose a height function ω10 on ∆1 that is identically 0 on abc such that T (ω1 ) =
T (ω10 ). Similarly we may choose ω20 on ∆1 that is identically 0 on abd such that T (ω2 ) = T (ω20 ). It follows
that ω10 > 0 and ω20 > 0 on all lattice points of ∆1 and ∆2 , respectively, outside of abcd. Since ω10 and ω20
agree on a and b, we can glue them together to obtain a height function ω 0 on ∆. Because of the positivity of
all coefficients away from abcd, we have that T (ω 0 ) is identical to T (ω), except that instead of the triangles
abc and abd we have the quadrilaterial abcd. We can the perform a pulling refinement [GOT18, §16.2] by
pulling at the lattice point c (or d) to obtain a regular triangulation T (ω 00 ) which has an edge from c to d,
and which is otherwise identical to T (ω 0 ). This regular triangulation is precisely the triangulation obtained
by performing our bistellar flip, thus completing the proof.

Proposition 3.7 helps us relate troplanarity of graphs of the same genus. For instance, we know that if
the seventh graph in Figure 8 were troplanar, then so would the sixth graph. Contrapositively, if we take it
as a given that the sixth graph is not troplanar, then we know the same is true of the seventh. We can also
use the proposition to bound the number of troplanar graphs by the number of 2-edge-connected troplanar
graphs.

Corollary 3.8. Let T (g) be the number of troplanar graphs of genus g, and let T (2) (g) be the number of
2-edge-connected troplanar graphs of genus g. Then T (g) ≤ 2g−1 T (2) (g).
Proof. Let G be a troplanar graph of genus G. Iteratively reduce all its bridges to obtain a 2-edge-connected
troplanar graph G0 . The order in which the bridge reductions are performed does not matter, so G0 is
well-defined.
We now ask how many distinct graphs could be bridge-reduced to the same graph G0 . Let us assume
that G0 is troplanar and 2-edge-connected. Let T be a regular unimodular triangulation of a lattice polygon
∆ giving rise to G0 . Viewing T as a subgraph whose vertices are ∆ ∩ Z2 , we then have that G0 is the dual
graph of the subgraph H of T induced by the interior lattice points of ∆. Every 2-edge-cut of G0 corresponds
to a bridge in H. Any graph G giving rise to G0 via a sequence of bridge reductions can be recoverd by
choosing some subset of the bridges in H, and reversing the bridge reduction surgery at each corresponding
2-edge-cut of G0 . Since H is a graph with g vertices, it has at most g − 1 bridges, so there are at most 2g−1
subsets of bridges of H. It follows that no more than 2g−1 graphs could give rise to G0 via bridge reductions.
Since every troplanar graph can be bridge reduced to a 2-edge-connected troplanar graph, we have
T (g) ≤ 2g−1 T (2) (g).
We close this section by presenting some results on the troplanar graphs for genus 6 and 7. To find all
such troplanar graphs, we must first determine the maximal nonhyperelliptic polygons of genus 6 and 7.

10
 Figure 16: The maximal nonhyperelliptic polygons of genus 6 and genus 7

Proposition 3.9. The maximal nonhyperelliptic polygons of genus 6 and genus 7 are those pictured in
Figure 16.
Proof. The main tool we use here is [Koe91, Lemma 2.2.13], which states that any maximal nonhyperelliptic
polygon ∆ is obtained by moving out the edges of its interior polygon ∆(1) . Thus we need only consider all
2-dimensional lattice polygons with 6 or 7 lattice points, and determine which can be pushed-out to obtain
another lattice polygon.
First we will argue that none of the interior polygons can be nonhyperelliptic themselves. Any nonhy-
perelliptic polygon has lattice width at least 3, and so any polygon with nonhyperelliptic interior polygon
has lattice width at least 5 by [CC12, Theorem 4]. But the only polygon of genus 6 with lattice width 5 is
the triangle with vertices at (0, 0), (5, 0), and (0, 5), whose interior polygon has genus 0; and as computed
in [Cas12, Table 1], no lattice polygon of genus 7 has lattice width 5. Thus, no maximal polygon of genus 6
or 7 has a nonhyperelliptic interior polygon.
Thus we need only consider candidate interior polygons that are hyperelliptic. We now use a classification
of hyperelliptic lattice polygons ∆ of genus g ≥ 0 presented in [Koe91] and reproduced in [Cas12, Theorem
10]. If the interior polygon has genus 0, it must either be a right trapezoid of height 1, or (in the case of 6
lattice points) a triangle with side lengths 2. These cases yield five polygons in Figure 16, namely the first
three polygons of genus 6 and the first two polygons of genus 7. If the interior polygon has genus 1, then
it must be one of the 16 genus 1 polygons pictured in [Cas12, Theorem 10(b)]. Two of these have 6 lattice
points, and four of these have 7 lattice points. It turns out every such polygon can be pushed out, yielding
the other two pictured polygons of genus 6, and the bottom row of polygons of genus 7.
Finally we deal with the case that the interior polygon has genus at least 2. As it is hyperelliptic, it must
have one of the forms classified in [Cas12, Theorem 10(c)]. For a genus 6 polygon, the interior polygon will
either have genus 2 or 3, and for genus 7, it will either have genus 2 or 3 or 4 (since any lattice polygon has
at least three boundary points). Running through the finitely many cases shows that there is a unique such
polygon with 6 or 7 lattice points that can be pushed out to a lattice polygon: it has 7 lattice points, and
yields the final polygon of genus 7 pictured on the right of the middle row in Figure 16.
We used TOPCOM [Ram02] to find all regular unimodular triangulations of these maximal nonhyper-
elliptic polygons, and then computed the resulting troplanar graphs. We would have also included the
graphs arising from the hyperelliptic polygons, although it turns out all such graphs also arose from non-
hyperelliptic polygons for g = 6 and g = 7. In the end we found that there are 151 troplanar graphs
of genus 6, and 672 troplanar graphs of genus 7. The complete lists of these graphs are available at
https://sites.williams.edu/10rem/supplemental-material/. Table 1 contains, for genus g from 2

11
to 7, the numbers of troplanar graphs T (g), of connected trivalent planar graphs P(g), and of connected
trivalent graphs G (g). (To our knowledge, the value of P(7) is not present in the literature, and so has
been omitted.) We remark that the sequence 2, 4, 13, 38, 151 does not match any sequence on the Online
Encyclopedia of Integer Sequences.

g T (g) P(g) G (g)

2 2 2 2
3 4 5 5
4 13 16 17
5 38 67 71
6 151 354 388
7 672 ? 2592

Table 1: The number of troplanar, planar, and general graphs that are connected and trivalent, by genus

Figure 17: All non-troplanar genus 6 graphs which are not ruled by any known criterion

To get a feel for how the results from earlier in this section rule out non-troplanar graphs, we review
some data on graphs of genus 6, all of which are illustrated in [Bal76]. Of the 388 connected trivalent graphs
of genus 6, a total of 151 are troplanar. This leaves 237 non-troplanar graphs, of which 34 are nonplanar.
Of the remaining 203 planar-but-not-troplanar graphs, 87 are sprawling and 63 are crowded (with 5 both
sprawling and crowded). This gives us 58 non-troplanar graphs not ruled out by any criterion known prior
to this paper. With our Theorem 3.4, we can rule out 19 more; and with Proposition 3.6 we rule out another
10. This leaves 29 graphs that are not troplanar, but are not ruled out by any known criterion; these graphs
are pictured in Figure 17.
We remark that none of the troplanar graphs of genus 6 have two adjacent vertices where each is also
incident to a bridge that is incident to a loop. If one could prove that a result akin to Corollary 3.5 forbidding
two loops in a row for graphs with g ≥ 6, this would rule out an additional 18 graphs of genus 6.

12
4 An upper bound on the number of troplanar graphs
Let g ≥ 2, and let us consider the proportion of connected trivalent planar graphs of genus g that are
troplanar; that is, let us consider T (g)/P(g). From Table 1, we have T (2)/P(2) = 2/2 = 1, T (3)/P(3) =
4/5 = 0.8, and so on; in other words, 100% of genus 2 planar graphs are troplanar, and 80% of genus 3
planar graphs are troplanar. We begin this section by proving that T (g)/P(g) tends to 0 as g tends to
infinity. We could also consider the number of troplanar graphs that are simple (that is, have no loops or
multiedges) compared to the number of simple connected planar graphs. Calling these numbers T s (g) and
P s (g), we will also prove that T s (g)/P s (g) tends to 0 as g tends to infinity. We will then develop a stronger
asymptotic upper bound on T (g), which will also imply that T (g)/P(g) tends to 0.

Theorem 4.1 (Theorem 5 in [BKLM07]). Let H be a fixed connected planar graph with one vertex of degree
1 and each other vertex of degree 3. Then there exists δ > 0 such that for all g, the probability that a random
connected trivalent planar graph of genus g contains fewer than δg copies of H is e−Ω(g) .
We note that the original result was stated in terms of n, the number of vertices of the graph, rather
than in terms of g. However, for a connected trivalent graph with n vertices we have n = 2g − 2, so the
result is easily translated into g. We also remark that the original result was stated for simple graphs; it is
easily generalized to multigraphs.
T (g) T s (g)
Theorem 4.2. We have limg→∞ P(g) = 0, and limg→∞ P s (g) = 0.
T (g)
Proof. We will prove that limg→∞ P(g) = 0; the other result follows from an identical argument, since the
graph H below is simple. Let H be the sprawling graph illustrated in Figure 18. Any trivalent graph
containing a copy of H must also be sprawling, with v serving as a disconnecting vertex. By Theorem 4.1,
there exists δ > 0 such that for all g, the probability that a random connected trivalent planar graph of
genus g contains fewer than δg copies of H is e−Ω(g) . For sufficiently large g, we have δg ≥ 1. This means
that the probability a graph does not contain any copy of H goes to 0 as g → ∞, which implies that the
T (g)
probability a graph is troplanar goes to 0 as g → ∞. We conclude that limg→∞ P(g) = 0.

Figure 18: The graph H used in the proof of Theorem 4.2

It follows from the above argument that most connected trivalent planar graphs (simple or otherwise)
are sprawling. Choosing different H’s can similarly show that most such graphs are crowded, and that most
such graphs are TIE-fighters, providing alternate proofs that most connected trivalent planar graphs are not
tropically planar. We remark that no graphs resulting from this argument are 2-edge-connected, and that it
is unknown how the 2-edge-connected counts T (2) (g) and P (2) (g) relate to one another.
The next result gives us an idea of how P(g) and P s (g) grow with g. Again, we have translated the
results from vertices to genus; we have also translated from labelled to unlabelled graphs, using the fact that
asymptotically 100% of cubic graphs have no nontrivial symmetries.

13
Theorem 4.3 (Theorems 1.1 and 1.3 from [NRR16]). Letting ∼ denote asymptotic equivalence, we have
7
P(g) ∼ c(2g − 2)− 2 γ 2g−2

where c ≈ 0.104705 and γ ≈ 3.985537; and we have

7
P s (g) ∼ d(2g − 2)− 2 ψ 2g−2

where d ≈ 0.030487 and ψ ≈ 3.132591.

Combined with Theorem 4.2, this immediately implies the following.
Corollary 4.4. We have T (g) = o((2g − 2)−7/2 ξ g ), where ξ = γ 2 ≈ 15.88451; and T s (g) = o((2g −
2)−7/2 ζ g ), where ζ = ψ 2 ≈ 9.813126.

We spend the remainder of the section proving a stronger asymptotic bound on T (g). Our main strategy
is to stratify T (g) by separating out the graphs coming from polygons of different lattice widths. Let T`lw (g)
denote the number of troplanar graphs arising from polygons of genus g with lattice width exactly `, and let
T≥`lw
(g) denote the number of troplanar graphs arising from polygons of genus g with lattice width at least `.
Since all polygons of genus g ≥ 2 have lattice width at least 2, we have T (g) ≤ T2lw (g) + T3lw (g) + T≥4
lw
(g),
where the inequality comes from the fact one graph may come from multiple polygons. We will separately
bound each of these three summands, thus giving an overall bound on T (g).
The easiest term to handle is T2lw (g). If a polygon has lattice width 2 and genus at least 2, it must
have a one-dimensional interior polygon and thus be hyperelliptic. Due to the work of [Cha13], [BJMS15,
§5], and [Mor17], the case for hyperelliptic polygons is understood quite well: for each genus g there are
2g−2 + 2b(g−2)/2c distinct skeletons that arise from some hyperelliptic polygon of genus g; see Section 5 for
more details. Thus we have an explicit formula for T2lw (g).
We now move on to bounding T3lw (g). Our starting point for this is the following result, which considers
a single polygon of lattice width 3.
√
Proposition 4.5. A nonhyperelliptic polygon of lattice width 3 gives rise to O(8g / g) troplanar graphs.
Proof. Let ∆ be a nonhyperelliptic polygon of lattice width 3. First we note that ∆ cannot be a standard
triangle, as the only standard triangle of lattice width 3 has genus 1 and is thus hyperelliptic. It follows from
[CC12, Theorem 4] that lw(∆(1) ) = lw(∆)−2 = 1. This means that ∆(1) has no interior lattice points, which
combined with its lattice width implies that up to equivalence, ∆(1) must be a right trapezoid of height one,
with vertices at (0, 0), (0, 1), (a, 0), and (1, b), where a ≥ b ≥ 1 and a + b + 2 = g [Koe91, Cas12].
√ √
To show that ∆ gives rise to O(8g / g) graphs, we will first show that it gives rise to O(4g / g) 2-edge-
connected troplanar graphs of genus g. A key fact is that if a triangulation T gives rise to a a 2-edge-connected
troplanar graph G, then G is determined by a very small amount of information in T , namely by the edges
connecting the interior lattice points of ∆. Every interior lattice point of ∆ in must contribute a bounded
face to the troplanar graph; since there are no bridges, for each pair of the g bounded faces of the graph,
they either share a common edge or are non-adjacent. Since two bounded faces share an edge if and only
an edge connects the corresponding interior points in the triangulation, we can draw the 2-edge-connected
troplanar graph with only this limited information.
Let T be a regular unimodular triangulation of ∆ giving rise to a 2-edge-connected troplanar graph G.
Let T 0 consist of all edges in T with both endpoints at interior lattice points of ∆. Letting ∂ ∆(1) denote

the boundary of ∆(1) , we claim that T 0 ∪ ∂ ∆(1) is a unimodular triangulation of ∆(1) , as illustrated in

Figure 19. Certainly it is a subdivison: any choice of non-crossing edges between the top and bottom rows
of a polygon of lattice width 1 gives a subdivision.
Suppose for the sake of contradiction that T 0 ∪ ∂ ∆(1) is not a unimodular triangulation. Then there

is a subpolygon Σ of ∆(1) present in T 0 ∪ ∂ ∆(1) that is not a unimodular triangle. The polygon Σ must

have two adjacent boundary points p and q such that the edge pq is not present in T ; otherwise Σ would be
a closed face in T , and T would not be a unimodular triangulation. Note that all edges in ∂(Σ) \ ∂(∆(1) )
are present in T , so pq ∈ ∂(∆(1) ). If p and q are in different rows, then either {p, q} = {(0, 0), (0, 1)} or
{p, q} = {(a, 0), (b, 1)}. In the first case, T must have either an edge from (1, 0) to (−1, 1) or from (1, 1) to

14
 T0 T 0 ∪ ∂(∆(1) )
T

Figure 19: A
triangulation T of a lattice width 3 polygon ∆; the subset T 0 ; and the resulting triangulation
T 0 ∪ ∂ ∆(1) of ∆(1)

(−1, 0), which then yields either an edge from (1, 0) to (0, 1) or from (1, 1) to (0, 0); this means that Σ is a
unimodular triangle, a contradiction. A similar contradiction arises if {p, q} = {(a, 0), (b, 1)}. Thus p and q
must be in the same row.
Because pq is not present in T , some e in T \ T 0 must pass through pq. It cannot be a nontrivial split,
since G is 2-edge-connected, so e must have endpoints r and s where r is an interior lattice point and s is a
boundary lattice point, as illustrated in Figure 20. Note that r must be on the opposite row from p and q.
At this point no edge in T can separate p from r: such an edge would have to cross e. Thus pr is an edge in
T . Similarly, qr is an edge in T . It follows that the triangle pqr appears in the subdivision T 0 ∪ ∂(∆(1) ) of
∆(1) . By Pick’s Theorem, pqr is unimodular. Since p and q are on the same row, they can only be common
vertices of one polygon in the
subdivision of ∆(1) , so Σ must be this unimodular triangle, a contradiction.
We conclude that T 0 ∪ ∂ ∆(1) is a unimodular triangulation.

s
e
p
q

Figure 20: The relative positions of p, q, r, and s. The polygon Σ is solid, and necessary edges in T are
dashed.

Ignoring for a moment whether (0, 0) connects to (0, 1) and whether (a, 0) connects to (b, 1) in T , this
means that the number of ways the upper and lower lattice points of ∆(1) can connect to one
another in
T is equal to the number of unimodular triangulations of ∆(1) . By [KZ03, §2], ∆(1) has a+ba unimodular
a+b g−2 g−2




triangulations. We have a + b = g − 2, so a = a ≤ b(g−2)/2c . This central binomial coefficient is
2g−2
asymptotic to √ .
π(g−2)/2
The only remaining data regarding the connections between the interior lattice points is which edges
from the boundary of ∆(1) are present in the triangulation T . Since ∆(1) has g boundary points, it also has
g boundary edges, so there are 2g ways to choose which are included and which are not. Multiplying this
by our binomial coefficient bound, we find that the number of ways the interior lattice points of ∆ could be
√
connected in order to yield a 2-edge-connected troplanar graph is O(4g / g).
By the same argument from Corollary 3.8, the total number of troplanar graphs arising from ∆ is at most
2g−1 times the number of 2-edge-connected graphs arising from ∆. We conclude the number of troplanar
√
graphs arising from ∆ is O(8g / g).
√
Corollary 4.6. We have T3lw (g) = O(8g · g).
Proof. As already noted, every nonhyperelliptic polygon of lattice width 3 and genus g has an interior polygon
equal to a right trapezoid of height 1 with g lattice points. There are O(g) such possible interior polygons,
and each has at most one associated maximal polygon. Every one of these maximal polygons contributes
√ √
O(8g / g) troplanar graphs, so in total we have T3lw (g) = O(8g · g).

15
 We now need to bound T≥4 lw
(g). It turns out that our argument will hold in general for T≥`+1
lw
(g) for any
` ≥ 3, so we work in that generality before specializing to ` = 3 to obtain our final result. We will find a
bound for the number of all lattice polygons of genus g; then we will find a bound for how many unimodular
triangulations a polygon of lattice width ` + 1 (or more) can have. Multiplying these together will give an
upper bound on T≥`+1
lw
(g).
For bounding the number of polygons, our starting point is the following result, which bounds the number
of polygons with at most a fixed area. The upper bound comes from [BV92], the lower bound from [Arn80].
Theorem 4.7. Let A be a fixed rational number, and let N (A) be the number of convex lattice polygons
distinct under lattice transformations with area less than A. Then there exist two positive constants C1 , C2
such that
C1 A1/3 ≤ log(N (A)) ≤ C2 A1/3
To convert this result from bounded area to bounded genus, we need to determine how many lattice
points are possible in a polygon of genus g, and then apply Theorem 2.1. We will use the following fact from
derived in [Cas12], presented as Equation (2) following the proof of their Theorem 1.
Proposition 4.8. Let ∆ be a nonhyperelliptic lattice polygon with v vertices and with r lattice boundary
points, and let r(1) be the number of lattice boundary boundary points of ∆(1) . Then

r ≤ r(1) + 12 − v.
Since v ≥ 3 and r(1) ≤ g, we immediately obtain the following corollary.
Corollary 4.9. If ∆ is nonhyperelliptic with r lattice boundary points and g interior lattice points, then
r ≤ g + 9.
By applying Pick’s Theorem and this corollary, we can then use Theorem 4.7 to obtain an upper bound
for the number of lattice polygons of genus g.
Lemma 4.10. There is a positive constant C2 such that the number of convex nonhyperelliptic lattice poly-
3g 1/3
gons of a genus g is at most eC2 ( 2 +9) .
Proof. If ∆ is a nonhyperelliptic polygon of genus g, it has at most g +9 lattice boundary points by Corollary
4.9. By Pick’s Theorem, the area of ∆ is r + g2 − 1, which is at most g + 9 + g2 − 1 = 3g 2 + 8. This means
that in order to bound the number of nonhyperelliptic polygons of genus g, it suffices to bound the number
of polygons with area less than 3g 3g
2 + 8 + 1 = 2 + 9. In the notation of Theorem 4.7, we are bounding
N 3g


2 + 9 . That theorem tells us there exists a positive constant C2 such that that
 
3g 3g
log(N + 9 ) ≤ C2 ( + 9)1/3 ,
2 2
which can be rewritten as  
3g 3g 1/3
N +9 ≤ eC2 ( 2 +9) .
2
3g 1/3
Thus the number of nonhyperelliptic polygons of genus g is bounded by eC2 ( 2 +9)

We now need to bound the number of unimodular triangulations a polygon of lattice width at least ` + 1
can have. To do this, we prove the following proposition, which can be viewed as a stronger version of
Corollary 4.9 for this class of polygons.

p width lw(∆) ≥ ` + 1 where ` ≥ 3, and let ∆

Proposition 4.11. Suppose that ∆ is a polygon with of lattice
have genus g and r lattice boundary points. Then r ≤ 2g
` + 4 g + 8/3 + 2

Proof. First we recall the bound lw(∆)2 ≤ 8Vol(∆)/3, proven in [FTM74] and also presented in [CC17,
Lemma 5.2(vi)]. By Pick’s theorem we know Vol(∆) = g + 2r − 1, and since ∆ cannot be hyperelliptic due
to its lattice width we know r ≤ g + 9 by Corollary 4.9. Combining these we find
8 r  8 g+9

32
lw(∆)2 ≤ 8Vol/3 = g+ −1 ≤ g+ − 1 = 4g + .
3 2 3 2 3

16
 p p
Square-rooting both sides yields lw(∆) ≤ 4g + 32/3 = 2 g + 8/3. We now use [Cas12, Theorem 8], which
states
2
r≤ · g + 2 · (lw(∆) + 1) .
lw(∆) − 1
p
Since lw(∆) ≥ ` + 1 and lw(∆) ≤ 2 g + 8/3, we have

2  p  2g p
r≤ · g + 2 · 2 g + 8/3 + 1 = + 4 g + 8/3 + 2,
(` + 1) − 1 `

as desired.
√
Proposition 4.12. A polygon of genus g and lattice width at least ` + 1 admits at most 2(3+ ` )g+4 g+8/3−1
2

unimodular triangulations
Proof. Let ∆ be a nonhyperelliptic polygon of genus g with r lattice boundary points and lattice width
3g+r−3
at least ` + 1. By [DLRS10, Theorem p 9.3.7], ∆ admits at most 2 unimodular triangulations. By
Proposition 4.11 we have r ≤ 2g
` + 4 g + 8/3 + 2. It follows that the number of unimodular triangulations
of ∆ is bounded by √
  √
3g+ 2g g+8/3+2 −3
` +4
= 2(3+ ` )g+4 g+8/3−1 .
2
2

Combined with our bound on the number of lattice polygons of genus g, this gives the following result.
√
(g) ≤ eC2 ( 2 +9) · 2(3+ ` )g+4 g+8/3−1 , where C2 is the constant from
3g 1/3 2
Corollary 4.13. We have T≥`+1
lw

Lemma 4.10.
We can now prove the following upper bound on T (g).
Theorem 4.14. We have  11g √ 
T (g) = O 2 3 +O( g) .

Proof. As already noted, we have

T (g) ≤ T2lw (g) + T3lw (g) + T≥4

lw
(g)
√
We know T2lw (g) = O(2g ), and byCorollary 4.6 we have T2lw (g) = O(8g / g). It follows from Corollary
11g √ 
4.13 when ` = 3 that T≥4
lw
(g) = O 2 3 +O( g) . The largest of the three bounds is the one coming from
T≥4lw
(g), and so this serves as our bound for T (g).

It follows that T (g) = o (211/3 + ε)g for any ε > 0, where 211/3 ≈ 12.699. This illustrates that Theorem

4.14 is indeed a stronger bound on T (g) than the result from Corollary 4.4. It also provides another argument

that most connected trivalent planar graphs are not tropical: by Theorem 4.3, P(g) = Ω (2g − 2)−7/2 ξ g
where ξ ≈ 15.88451. Due to this exponential base being larger than 211/3 + ε, the ratio T (g)/P(g) rapidly
goes to 0 as g increases.
We close this section by discussing a few ways in which our upper bound might be improved. One strategy
could be to stratify further by lattice width. Fix ` ≥ 3, and consider the bound

T (g) ≤ T2lw (g) + · · · + T`lw (g) + T≥`+1

lw
(g).

We know by Corollary 4.13 that  √ 

(g) = O 2(3+ ` )g+O( g) .
2
T≥`+1
lw

If we can effectively bound Tilw (g) for i ≤ `, this could lead to an improvement on Theorem 4.14. So far we
have bounds on T2lw (g) and T3lw (g), so T4lw (g) would be the next step. We remark that this strategy will
never lead to a stronger bound than O(2(3+ε)g ) on T (g).

17
 Another possible direction for future improvement comes from the observation that the bound from
Proposition 4.12 is a bound on all unimodular triangulations, when we only need to consider those uni-
modular triangulations that are regular. Numerics such as those computed in [KZ03] suggest that regular
triangulations are much rarer than non-regular triangulations for large polygons. A precise enough result to
this effect could lower the number of triangulations we consider.
Even restricting to regular triangulations, we note that many triangulations can give the same graph, even
though our bound from Corollary 4.13 comes from bounding the total number of triangulations. For instance,
as computed in [BJMS15], the unique maximal nonhyperelliptic polygon of genus 3 admits 1278 regular
unimodular triangulations up to symmetry, but only yields four distinct troplanar graphs. Unfortunately,
we cannot in general say that each troplanar graph arises from many different triangulations: the graph
of genus 4 illustrated in Figure 6 only arises from a single triangulation of a single polygon, as shown in
[BJMS15].
We also remark on one way in which our argument is already essentially optimal. Although it is a less
dramatic contributor to our bound, we might wonder if we are significantly overestimating the number of
polygons of genus g that we must consider. The reader may notice that the bound from Lemma 4.10 is
for general convex lattice polygons, while we only need to work with maximal polygons. The following
arguments show that in fact this will not meaningfully change our upper bound from Lemma 4.10.
Note that if a polygon is nonmaximal, then at least one of its edges has lattice length 1. This is because
any nonmaximal polygon can be constructed by taking a maximal polygon with the same interior polygon,
removing carefully chosen boundary points, and taking the convex hull of the remaining points; assuming the
interior polygon has not been changed, this always produces an edge of lattice length 1. Contrapositively,
if all sides of a polygon have lattice length 2 or more, then that polygon is maximal. Let ∆ be a lattice
polygon, and let ∆0 be 2∆, the polygon obtained by scaling ∆ by a factor of 2. In going from ∆ to ∆0 , the
perimeter increases by a factor of 2, and the area increases by a factor of 4.
Lemma 4.15. Let ∆ and ∆0 have g and g 0 interior lattice points and r and r0 boundary lattice points,
respectively. Then g 0 = 4g + r − 3. Moreover, if g ≥ 1, then g 0 ≤ 6g + 4.
0
Proof. By Pick’s theorem, the area of ∆ is 2r + g − 1 and the area of ∆0 is r2 + g 0 − 1, We know that the
0
area of ∆0 is four times that of ∆, so 2r + 4g − 4 = r2 + g 0 − 1. We also know that r0 = 2r, since the number
of boundary points is equal to the (lattice) perimeter, giving us 2r + 4g − 4 = r + g 0 − 1. Solving for g 0 gives
g 0 = 4g + r − 3.
For the inequality on g 0 , we will use the fact that r ≤ 2g + 7 for any lattice polygon of genus at least 1
[Sco76]. Since g 0 = 4g + r − 3, we have g 0 ≤ 4g + (2g + 7) − 3 = 6g + 4, as claimed.

Proposition 4.16. Let g ≥ 1. If there are N lattice polygons of positive genus at most g, then there are at
least N maximal lattice polygons of genus at most 6g + 4.
Proof. Assume there are N (distinct) lattice polygons of positive genus at most g. Mapping each polygon
∆ to 2∆ gives a collection of N distinct maximal lattice polygons, since each side of 2∆ has length at least
2. By the previous lemma, each of these polygons has genus at most 6g + 4, proving our claim.

We now use the lower bound from Theorem 4.7, which says that there is a positive constant C1 such that
C1 A1/3 ≤ log(N (A)), where N (A) is the number of convex lattice polygons with area less than A ∈ Q+ .
Note that a lattice polygon of genus g has area at least g + 23 − 1 = g + 21 by Pick’s Theorem. This means
that the number of polygons of genus at most g is an upper bound for N (g + 1), the number of polygons with
area less than g + 1. Thus any lower bound on N (g + 1) is also a lower bound on the number of polygons of
1/3
genus at most g. By a similar argument to that presented in Lemma 4.10, we have eC1 (g+1) ≤ N (g + 1),
1/3
so there are at least eC1 (g+1) polygons of genus at most g. Only a few of these polygons can have genus
0: by the classification result in [Koe91, Cas12], only quadratically many genus 0 polygons have a fixed area
A, so only cubically many genus 0 polygons have area at most A. This is dwarfed by the exponential in
0 1/3
g 1/3 , so we can replace C1 with another positive constant C10 to conclude that there are at least eC1 (g+1)
polygons of positive genus at most g (assuming that g ≥ 1).
0
Solving g 0 ≤ 6g + 4 for g gives g ≥ g 6−4 , yielding the following bound.

18
 0 1/3
Corollary 4.17. There exists a positive constant C10 such that for all g ≥ 10, there are at least eC1 ((g+2)/6)
maximal polygons of genus at most g.
0 1/3
Proof. Our assumption on g guarantees that (g − 4)/6 ≥ 1. There are at least eC1 (((g−4)/6)+1) =
0 1/3
eC1 ((g+2)/6) lattice polygons of positive genus at most (g − 4)/6. For each such polygon ∆, the maxi-
0 1/3
mal polygon 2∆ has genus at most g. It follows that there are at least eC1 ((g+2)/6) maximal polygons of
genus at most g.

5 A lower bound on the number of troplanar graphs

As mentioned briefly in the previous section, hyperelliptic polygons of genus g give rise to 2g−2 + 2b(g−2)/2c
distinct skeletons. To see where this formula comes from, we recall the following argument from [BJMS15,
§6]. All graphs that arise from hyperelliptic polygons are chains, meaning that they have the structure of
a sequential chain of loops with two adjacent chains either sharing a common edge or being joined by a
bridge. The three chains of genus 3 are illustrated in Figure 21, along with subdivisions of a hyperelliptic
polygon that give rise to them. One can specify a chain by a binary string of length g − 1, describing what
the sequence of bridges and shared edges is; for the illustrated graphs, these strings would be 00, 01, and 11,
where 0 is a shared edge and 1 is a bridge. Two strings give the same graph if and only if they are reverses of
one another, like 01 and 10. The number of binary strings of length g − 1, up to this reversing equivalence,
is 2g−2 + 2b(g−2)/2c . This gives us a lower bound on T (g), and we may write T (g) = Ω(2g ).

Figure 21: The three chains of genus three, and triangulations of a hyperelliptic polygon giving rise to them

Our goal for this section is to provide a better lower bound on T (g) by constructing a family of graphs
that is asymptotically larger than the family of chains. To do so, we will construct many unimodular
triangulations of a polygon of genus g, argue that these triangulations are regular, and prove that each
triangulations gives rise to different troplanar graph. These are in general challenging endeavors which are
much simpler in the hyperelliptic case. First, all unimodular triangulations of hyperelliptic polygons are
regular [KZ03, Proposition 3.4]. Second, determining whether two chains are isomorphic is relatively simple:
they must have the same pattern of shared edges and bridges, up to perhaps flipping the graph around. Our
family of troplanar graphs will be constructed so that each pair of graphs can be similarly verified to be
non-isomorphic. We begin with the following result.
Proposition 5.1. Suppose we have two trivalent graphs G and H of genus g that are of the following forms:
c

a G1 G2 Gk−1 Gk b
pL pR pL pR pL pR pL pR
1 1 2 2 k−1 k−1 k k

c0

a0 H1 H2 Hm−1 L
Hm b0
qL qR qL qR qL qR qm R
qm
1 1 2 2 m−1 m−1

d0

where each box contains a positive-genus 2-edge-connected component Gi of G or Hi of H, respectively, with

exactly two 2-valent vertices where the bridges connect; call these vertices pL R L R
i and pi for Gi , and qi and qi

19
for Hi . Then G is isomorphic to H if and only if k = m and for each i, Gi is isomorphic to Hi via an
isomorphism sending pL L R R
i to qi and pi to qi .

Proof. If k = m and Gi is isomorphic to Hi for all i via an isomorphism sending pL L R R

i to qi and pi to qi , then
there is a natural isomorphism between G and H: simply use the isomorphisms on the 2-edge-connected
components, and send a, b, c, and d to a0 , b0 , c0 , and d0 , respectively.
Now assume G is isomorphic to H. Since G has k + 1 bridges and H has m + 1 bridges, we must have
k = m since isomorphism preserves the number of bridges. Let ϕ : V (G) → V (H) be a graph isomorphism.
Since a and a0 are the only vertices incident to a loop, we must have ϕ(a) = a0 . Since a is only adjacent to
0
pL L L L
1 and a is only adjacent to q1 , we must have ϕ(p1 ) = q1 . Any isomorphism of graphs will map 2-edge-
connected components to 2-edge-connected components, and since one vertex in G1 is mapped to one vertex
in H1 , ϕ restricted to V (G1 ) must give an isomorphism from G1 to H1 . Only two vertices in G1 are incident
to a bridge in G, namely pL R L R L
1 and p1 ; the same is true for H1 in H, namely q1 and q1 . Since ϕ(p1 ) = q1 ,
L
R R L L
we must have ϕ(p1 ) = q1 . Thus, G1 and H1 are isomorphic via an isomorphism that maps p1 to q1 and
pR R R L
1 to q1 . Since all vertices incident to p1 have been accounted for except p2 , and since the same holds for
R L L L
q1 except for q2 , we must have ϕ(p2 ) = q2 . We may then apply an identical argument to establish the
desired isomorphism from G2 to H2 , and so on, all the way up to the fact that pR R
k must be sent to qk . This
completes the proof.
We will construct a number of regular triangulations which give rise to graphs of the form considered in
||
Proposition 5.1. For even g, let Pg denote the parallelogram of genus g with vertices at (0, 3), (1, 0), (g/2, 3),
||
and ((g + 2)/2, 0), as pictured in Figure 22. We will construct regular triangulations of Pg by tiling it with
|| || ||
triangulated copies of P2 , P4 , and P6 . We will then slightly modify the polygon so that the troplanar
graphs obtained from our triangulations have the forms prescribed by Proposition 5.1.

(0, 3) (g/2, 3)

(1, 0) ((g + 2)/2, 0)

||
Figure 22: The parallelogram Pg
|| || ||
We will refer to the triangulated copies of P2 (respectively of P4 and P6 ) as tiles of genus 2 (respectively
tiles of genus 4 and tiles of genus 6). To start out, we will use only one tile of genus 2, namely the one
illustrated in Figure 23, which also pictures a dual tropical curve and the corresponding troplanar graph.
||
This is not the only troplanar graph that can be obtained from P2 ; however, it is the only bridge-less one,
which is important if we want to apply Proposition 5.1. We have also marked two points on the graph as L
and R: these are the points where the bridges would attach to this 2-edge-connected component if the tile
||
appeared in the middle of a triangulation of a larger Pg .

L R

Figure 23: The tile of genus 2, a dual tropical curve, and the troplanar graph with two marked points

We will start out with 8 tiles of genus 4, as illustrated in Figure 24. Although the boxed groups of graphs
would be isomorphic without the marked points, the marked points make them distinct. We will also begin
with 49 tiles of genus 6, presented in Appendix A. It is important to verify that all the tiles we use are
|| ||
regular triangulations. Using TOPCOM [Ram02], we searched for all non-regular triangulations of P2 , P4 ,
||
and P6 ; it turns out these polygons have no non-regular triangulations, so all our tiles are safe to use.

20
 L L
R L R L
L R L
L R
R
R
L R
R

Figure 24: The first eight tiles of genus 4

||
An example of a tiling of P14 is presented in the top left Figure 25. This is a regular triangulation,
since patching regular triangulations along edges of lattice length 1 preserves regularity, as noted in [KZ03,
Proposition 3.4]. The troplanar graph obtained is also illustrated. Although it is not quite of the form
prescribed by Proposition 5.1, a small modification of the polygon can fix this. In particular, we can expand
our polygon to the quadrilateral of genus 17 pictured in the bottom left of Figure 25, and triangulating it
as shown adds on the left loop and the right genus-2 structure.

L L

R L R

L L
L
R
R L R

||
Figure 25: A tiling of P14 , and the corresponding troplanar graph; followed by a tiling of Qodd
7

For an integer n let Qodd

n be the trapezoid with vertices at (0, 3), (2, 0), (n + 2, 3), and (n + 3, 0). This
||
polygon has genus 2n + 3, and consists of a copy of P2n with two polygons glued on, one of genus 1 and one
of genus 2. Similarly, let Qeven
n be the pentagon with vertices at (0, 1), (0, 3), (2, 0), (n + 2, 3), and (n + 3, 0).
These polygons are illustrated in Figure 26.

(0, 3) (n + 2, 3) (0, 3) (n + 2, 3)

(0, 1)
(n + 3, 0) (n + 3, 0)
(2, 0) (2, 0)

Figure 26: The polygons Qodd

n and Qeven
n , partially triangulated

21
Proposition 5.2. Let T and T 0 be triangulations of the same polygon, either Qodd n or Qeven
n , constructed by
||
completing the partial triangulations in Figure 26 by tiling the remaining copy of P2n with some sequences
of our 1 tile of genus 2, our 8 tiles of genus 4, and our 49 tiles of genus 6. Then T and T 0 are regular, and
the troplanar graphs arising from them are isomorphic if and only if they were constructed using the exact
same sequence of tiles.
Proof. Assume for the moment that we are triangulating the polygon Qodd n . Let T be constructed using
the sequence of tiles (T1 , · · · , Tk ), and let T 0 be constructed using the sequence of tiles (T10 , · · · , Tm 0
). Our
0
triangulations T and T are both regular, since they arise from regular triangulations glued along shared
edges of lattice length 1.
Let G and H be the troplanar graphs arising from T and T 0 . Certainly if they were constructed from
the same sequence of tiles, we have that G is isomorphic to H. Assume now that G is isomorphic to H.
By construction, both G and H are of the form in Proposition 5.1, where G1 , · · · , Gk are the graphs arising
from the tiles T1 , · · · , Tk , and H1 , · · · , Hm are the graphs arising from the tiles T10 , · · · , Tk0 . By Proposition
5.1, k = m and we have that for all i, each Gi is isomorphic to Hi via an isomorphism respected the left and
right marked points. Thus for all i, Ti and Ti0 give rise to isomorphic graphs, including the marked points.
By construction, each tile gives rise to a different marked graph, so Ti = Ti0 for all i. We conclude that G
and H were constructed from the exact same sequence of tiles.
This argument carries over to the polygon Qeven n , with the additional footnote that G1 and H1 are both
a biedge coming from the lattice point (1, 2) in Qeven n .

Thus in order to find a lower bound on the number of troplanar graphs of genus g + 3 or g + 4, we can
||
count the number of ways to tile the parallelogram Pg with our tiles of genus 2, 4, and 6. Before we do this,
however, we will argue that we can actually include even more tiles than presented so far, in particular ones
that yield graphs with bridges.
We will add one additional tile of genus 2, and five additional tiles of genus 4; these are pictured in
Figure 27, along with their marked troplanar graphs. We will also add 26 tiles of genus 6, illustrated in
Appendix A. We have selected these tiles so that (when taking markings into consideration) no two give the
same ordered pair of graphs, and so that no 2-edge-connected component they contribute is available from
a single tile. (This is not immediately obvious for the final tile of genus 4 pictured; however, the location of
the bridge relative to the marked points cannot be obtained from our tiles of genus 2.)

L R
L R R L L R R L
L R

Figure 27: The tiles of genus 2 and 4 giving graphs with bridges

Proposition 5.3. The result from Proposition 5.2 still holds with our additional tiles.
Proof. We will assume we are triangulating the polygon Qodd n ; a similar argument holds for Qn
even
.
0 0
Let T , T , G, and H be as in Proposition 5.2, where the sequences of tiles for T and T are (T1 , · · · , Tr )
and (T10 , · · · , Ts0 ). It is still the case that G and H are of the form prescribed in Proposition 5.1; however,
it might now be that a tile Ti (or Ti0 ) contributes more than one 2-edge-connected component to G (or H).
That is, we might have r < k or s < m.
Assume that G is isomorphic to H, and suppose for the sake of contradiction that (T1 , · · · , Tr ) 6=
(T10 , · · · , Ts0 ). Since the two sequences must contribute a total of genus 2n, they must differ in some tth entry
where 1 ≤ t ≤ min{r, s}. Let i be the first index for which Ti 6= Ti0 . Since (T1 , · · · , Ti−1 ) = (T10 , · · · , Ti−1
0
),

22
isomorphism between G and H is preserved if we contract all portions of G and H arising from the first i − 1
tiles; thus, we may assume without loss of generality that i = 1.
Since G is isomorphic to H, we know that G1 is isomorphic to H1 and G2 is isomorphic to H2 , with the
isomorphisms respecting the left and right marked points. The graphs G1 and H1 must come from T1 and
T10 . Since T1 6= T10 , at least one of these tiles, say T1 , must contribute another 2-edge-connected component
to G; thus G2 also comes from T1 . However, by construction no tile contributing two 2-edge-connected
components contributes a 2-edge-connected component that is available from a tile contributing only one 2-
edge-connected component. It follows that T10 must also contribute H2 . This is all T1 and T10 can contribute,
since no tile contributes more than two 2-edge-connected components. However, no distinct pair of our tiles
T1 and T10 give the same ordered pair of marked 2-edge-connected components, a contradiction. We conclude
that (T1 , · · · , Tr ) = (T10 , · · · , Ts0 ).

||
Let an denote the number of ways to tile the parallelogram P2n with our tiles. Our propositions imply
that each of the an tilings gives a distinct troplanar graph of genus 2n + 3, and a distinct troplanar graph
of genus 2n + 4. Thus, we have T (g) ≥ ab(g−3)/2c . To obtain an explicit lower bound on T (g), we will
determine a formula for an .
||
Proposition 5.4. Let an be the number of ways to tile the parallelogram P2n of genus g = 2n with our tiles
of genus 2, 4 and 6. Then an = Ω(αn ), where α ≈ 6.1233 is the unique real root of x3 − 2x2 − 13x − 75.
|| ||
Proof. Any tiling of P2n will either begin with a tile of genus 2, followed by a tiling of P2(n−1) ; or with a tile
|| ||
of genus 4, followed by a tiling of P2(n−2) ; or with a tile of genus 6, followed by a tiling of P2(n−3) . It follows
that an satisfies the recurrence relation

an = 2an−1 + 13an−2 + 75an−3

for n ≥ 3. The initial conditions are a0 = 1 (there is one way to tile the empty parallelogram, namely not
to tile anything), a1 = 2 (since there are two tiles of genus 2), and a2 = 17 (13 from the tiles of genus 4, and
4 from the 4 ways to place two tiles of genus 2).
The solution to this recurrence relation can be found by studying the roots of the characteristic polynomial
x3 − 2x2 − 13x − 75. This polynomial has three distinct roots: a real root α ≈ 6.1233, and two complex
conjugate roots β = reiθ and β = re−iθ , where r ≈ 3.4998 and θ ≈ 2.2007 (in radians). Any (complex)
solution to the recurrence relation has the form
n
an = Aαn + Bβ n + Cβ ,

where A, B, C ∈ C. Our initial conditions imply

A + B + C = 1,

Aα + Bβ + Cβ = 2,
2
Aα2 + Bβ 2 + Cβ = 17,
which we can rewrite as
1 1 1
    
A 1
α β β  B  =  2  .
2
α2 β2 β C 17
Numerically solving this system of equations using Mathematica [Inc], we find A ≈ 0.49999, B ≈ 0.25001 +
0.00543i, and C = B. We can then rewrite our solution as

an = Aαn + 2Drn cos(nθ − δ),

A n
where D = |B| and δ = arg(B). Since A > 0 and α > r, we have that |2Drn cos(nθ − δ)| ≤ 2α for
sufficiently large n. This implies that
A n
an = Aαn + 2Drn cos(nθ − δ) ≥ α
2

23
 A
for sufficiently large n. Since 2 > 0, this means that an = Ω(αn ).

This allows us to deduce the following asymptotic lower bound on T (g).

√
Corollary 5.5. We have T (g) = Ω(γ g ) as g → ∞, where γ = α ≈ 2.47 is the the square-root of the
unique real root of x3 − 2x2 − 13x − 75.
Proof. By Propositions 5.2 and 5.3, we have have T (g) ≥ ab(g−3)/2c for all g ≥ 5. Since an = Ω(αn )
√ g
by Proposition 5.4, we have ab(g−3)/2c = Ω(αb(g−3)/2c ) = Ω(αg/2 ) = Ω( α ). We conclude that T (g) =
√ g
Ω( α ).
We close our paper with several possible directions for future research. Our upper and lower bounds are
roughly exponential in g, with bases around γ ≈ 2.47 and 211/3 ≈ 12.699, respectively; closing a gap between
these bases is a natural goal for future work. For a better lower bound, the methods of Section 5 could be
pushed further, perhaps by considering tiles of higher genus. One approach for a better upper bound, in the
spirit of Proposition 4.5, could be to bound T (2) (g), the number of 2-edge-connected troplanar graphs. If
one proved a bound of the form T (2) (g) = O(bg ), then Corollary 3.8 would imply T (g) = O((2b)g ). Thus
to improve Theorem 4.14, it would suffice to show T (2) (g) = O(bg ) for some b < 211/3 /2 ≈ 6.3496.
In general there is no easily implemented single criterion (or list of such criteria) for determining whether
a graph is troplanar, and finding new criteria could be helpful in furthering our understanding of such graphs.
For instance, although many results (such as Proposition 3.1 and Theorem 3.4) forbid certain structures that
cannot be dual to general unimodular triangulations, no existing work has specifically used the regularity
of triangulations to forbid certain structures. In fact, it is unknown whether considering dual graphs of
nonregular triangulations yields any non-troplanar graphs. There is also no known example of a 2-edge-
connected trivalent non-troplanar graph that is neither nonplanar nor crowded. Finding such an example
could point to other general obstructions to troplanarity.
There are other reasonable families of graphs one could study that get at the notion of being “tropically
planar.” In the notation of [BJMS15], the moduli space of all metric graphs of genus g arising from smooth
tropical plane curves is written Mplanar
g . Asking if a graph G is tropically planar, then, is asking whether G
appears (for at least one choice of edge lengths) in Mplanar
g . However, one could also study trop(Mnd g ), the
tropicalization of the moduli space of nondegenerate curves studied in [CV09]. This is the space of all metric
graphs of genus g that arise as the Berkovich skeleton of a nondegenerate curve, and strictly contains Mplanar
g
for g ≥ 3. There are certainly combinatorial types of graphs that appear in trop(Mnd g ) that do not appear
in Mplanar
g ; for instance, all graphs of genus 3 and 4 appear in trop(Mnd nd
3 ) and trop(M4 ), respectively, even
though not all graphs of genus 3 and 4 are tropically planar. Even for g = 5, determining which combinatorial
types of graphs appear in trop(Mnd g ) appears to be an open question.
There are also alternate definitions of what is meant by a “tropical plane curve.” Instead of considering
tropical curves that are a subset of R2 , one could more generally consider tropical curves that are subsets
of 2-dimensional tropical linear spaces embedded in some Rn . The authors of [MAH19] take this approach,
and show that every metric graph of genus 3 (with a low-dimensional set of exceptions) arises in such a
2-dimensional tropical linear space in Rn for n ≤ 5; this includes metric graphs whose underlying graph is
the non-troplanar graph of genus 3. A natural question to ask is which graphs of genus g appear in tropical
curves embedded in such tropical linear spaces for g ≥ 4.

A Tiles of genus 6
In this appendix we present the tiles of genus 6 used in our construction of troplanar graphs in Section
5. The 49 tiles yielding bridgeless graphs are pictured in Figure 28. The 26 tiles yielding graphs with
bridges are pictured in Figure 29. In both figures, we box together isomorphism classes of graphs, which are
distinguished by the marked L and R points.

24
References
[Arn80] V. I. Arnold. Statistics of integral convex polygons. Funktsional. Anal. i Prilozhen., 14(2):1–3,
1980.
[Bak08] Matthew Baker. Specialization of linear systems from curves to graphs. Algebra Number Theory,
2(6):613–653, 2008. With an appendix by Brian Conrad.

[Bal76] A.T. Balaban. Chemical applications of graph theory. Academic Press, 1976.
[BJMS15] Sarah Brodsky, Michael Joswig, Ralph Morrison, and Bernd Sturmfels. Moduli of tropical plane
curves. Res. Math. Sci., 2:Art. 4, 31, 2015.
[BKLM07] Manuel Bodirsky, Mihyun Kang, Mike Löffler, and Colin McDiarmid. Random cubic planar
graphs. Random Structures Algorithms, 30(1-2):78–94, 2007.
[BV92] I. Bárány and A. M. Vershik. On the number of convex lattice polytopes. Geom. Funct. Anal.,
2(4):381–393, 1992.
[Cas12] Wouter Castryck. Moving out the edges of a lattice polygon. Discrete & Computational Geom-
etry, 47(3):496–518, Apr 2012.

[CC12] Wouter Castryck and Filip Cools. Newton polygons and curve gonalities. J. Algebraic Combin.,
35(3):345–366, 2012.
[CC17] Wouter Castryck and Filip Cools. Linear pencils encoded in the Newton polygon. Int. Math.
Res. Not. IMRN, (10):2998–3049, 2017.

[CDMY16] Dustin Cartwright, Andrew Dudzik, Madhusudan Manjunath, and Yuan Yao. Embeddings and
immersions of tropical curves. Collect. Math., 67(1):1–19, 2016.
[Cha13] Melody Chan. Tropical hyperelliptic curves. J. Algebraic Combin., 37(2):331–359, 2013.
[CV09] Wouter Castryck and John Voight. On nondegeneracy of curves. Algebra Number Theory,
3(3):255–281, 2009.
[DLRS10] Jesús A. De Loera, Jorg Rambau, and Francisco Santos. Triangulations, volume 25 of Algorithms
and Computation in Mathematics. Springer-Verlag, Berlin, 2010.
[FTM74] L. Fejes Tóth and E. Makai, Jr. On the thinnest non-separable lattice of convex plates. Stud.
Sci. Math. Hungar., 9:191–193 (1975), 1974.

[GOT18] Jacob E. Goodman, Joseph O’Rourke, and Csaba D. Tóth, editors. Handbook of discrete and
computational geometry. Discrete Mathematics and its Applications (Boca Raton). CRC Press,
Boca Raton, FL, 2018. Third edition of [ MR1730156].
[Inc] Wolfram Research, Inc. Mathematica, Version 11.3. Champaign, IL, 2018.

[Koe91] R. Koelman. The number of moduli of families of curves on a toric surface. PhD thesis,
Katholieke Universiteit de Nijmegen, 1991.
[KZ03] Volker Kaibel and Günter M. Ziegler. Counting lattice triangulations. In Surveys in combi-
natorics, 2003 (Bangor), volume 307 of London Math. Soc. Lecture Note Ser., pages 277–307.
Cambridge Univ. Press, Cambridge, 2003.

[MAH19] Yue Ren Ilya Tyomkin Marvin Anas Hahn, Hannah Markwig. Tropicalized quartics and canonical
embeddings for tropical curves of genus 3. Int. Math. Res. Not. IMRN, 2019.
[Mor17] Ralph Morrison. Tropical hyperelliptic curves in the plane, 2017.

25
[MS15] Diane Maclagan and Bernd Sturmfels. Introduction to tropical geometry, volume 161 of Graduate
Studies in Mathematics. American Mathematical Society, Providence, RI, 2015.
[NRR16] Marc Noy, Clement Requile, and Juanjo Rue. Random cubic planar graphs revisited. Electronic
notes in discrete mathematics, 54:211–216, 2016.

[Ram02] Jörg Rambau. TOPCOM: Triangulations of point configurations and oriented matroids. In
Arjeh M. Cohen, Xiao-Shan Gao, and Nobuki Takayama, editors, Mathematical Software—ICMS
2002, pages 330–340. World Scientific, 2002.
[Sco76] P. R. Scott. On convex lattice polygons. Bull. Austral. Math. Soc., 15(3):395–399, 1976.

26
 R R L R L
L R R L
L
L R L R L R

R L
L R L
L R L R R
L R R L
L R

L R L
R L R L
L L R R L
R L R
R

R L
L R R L
L R L R R L
L R
R L

R L R R L
L R
L R L R L
R L R L R L

R R L
R L
L
R
R L L L R
L R R L

Figure 28: The tiles of genus 6 giving graphs without bridges

27
L R R L R R L
L
L R L R

L R L R L
R
L R L R
L R

L R R L L R R L R L
L R

L R R L L R R L

L R R L L R

R L

Figure 29: The tiles of genus 6 giving graphs with bridges

28